Urban densification change detection for small-cell site re-assessment
New construction invalidates small-cell propagation models before a single antenna is installed. Stereo optical change detection between baseline and current epochs identifies new obstructions at building level, feeding updated clutter layers directly into 5G re-runs.
Sensors
- Pleiades Neo: 30 cm native panchromatic resolution, stereo and tri-stereo collection in a single pass. Digital surface models derived from stereo pairs achieve 50 cm horizontal and roughly 1 m vertical RMSE in urban areas under controlled GCP conditions. Revisit up to twice daily at mid-latitudes.
- Maxar WorldView Legion: 30 cm panchromatic, up to six satellites providing revisit of less than 30 minutes over major cities. Stereo DSM generation workflow equivalent to WorldView-3, which has a published 3D accuracy specification of 1 m CE90 without ground control.
- Sentinel-2 MSI: 10 m multispectral bands (visible and near-infrared), 5-day revisit at the equator with two satellites. Insufficient resolution for individual building footprints in dense cities, but useful as a free, frequent screening layer to flag construction-activity zones before tasking commercial stereo.
- OpenStreetMap building change logs: Timestamped polygon edits in the OSM history API provide a crowd-sourced prior on new building footprints. Coverage and latency are inconsistent by city, but OSM diffs can narrow the search area and reduce commercial tasking costs.
Why a six-month-old clutter layer is already wrong
A small-cell propagation model is only as good as the clutter layer it runs on. In a fast-developing city, a building that did not exist when the radio-frequency planning team ran their initial model can place a hard shadow across the intended coverage footprint. The problem is not hypothetical: construction timelines in high-growth urban corridors routinely produce multi-storey reinforced-concrete structures within the planning and permitting window between initial site selection and antenna installation.
The conventional response is a resurvey: a drive-test team or a drone operator revisits the candidate site. That costs time and money, and it scales badly when a network operator is assessing dozens of candidate sites across a city simultaneously. Change detection from archived and freshly tasked stereo imagery does the same job faster and at lower marginal cost per site, with the added advantage of a documented audit trail.
What stereo imagery actually measures, and where it falls short
A stereo pair from Pleiades Neo or WorldView Legion produces a digital surface model, not a bare-earth terrain model. The DSM captures the top of everything: rooftops, plant rooms, parapet walls, cranes, and trees. Subtracting a baseline DSM from a current DSM gives a height-change raster. Positive values indicate something new above the old surface; negative values indicate removal or demolition.
The vertical accuracy of commercially derived urban DSMs is typically cited at 1 to 2 m RMSE without ground control points, improving to around 0.5 m with a sparse GCP network. That means a two-storey extension (roughly 6 m) is detectable with high confidence, but a single-storey plant-room addition (3 m) sits close to the noise floor and should be treated as probable rather than certain. Horizontal positional accuracy at 30 cm native resolution is generally better than 1 m CE90 with rational polynomial coefficients, which is sufficient for footprint attribution to a specific parcel.
Cloud cover is the most significant operational constraint. A stereo collection requires a cloud-free window over both passes of the same scene. In tropical or monsoonal cities, this can delay acquisition by weeks. Sentinel-2 time series can bridge the gap by flagging construction activity through spectral change (bare soil and concrete have distinct near-infrared signatures relative to vegetation or sealed surfaces), but Sentinel-2 cannot resolve individual building outlines in dense urban fabric.
Separating new construction from cranes and demolition
A raw height-change map conflates three distinct phenomena that have very different implications for propagation planning: permanent new structures, temporary obstructions such as tower cranes, and demolition voids. Treating a crane as a permanent building would trigger an unnecessary model re-run; missing a new eight-storey residential block would leave the propagation model wrong.
Attribution relies on multi-temporal stacking. A crane appears as a tall, narrow positive anomaly that moves laterally between acquisitions taken weeks apart. A new building grows in height monotonically across a sequence of images and eventually stabilises at a finished floor plate. Demolition produces a persistent negative change that does not recover. Running three or four acquisitions across a six-month window, rather than a single before-and-after pair, gives enough temporal signal to classify each change polygon with reasonable confidence. Morphological filters on the height-change raster, combined with building-footprint priors from cadastral data or OSM, further reduce false positives from rooftop equipment swaps.
From change polygons to updated clutter layers
The output of the detection step is a set of change polygons, each attributed with height delta, change class (new build, demolition, temporary, ambiguous), and a confidence score. These polygons are not the final deliverable for a radio-frequency engineer: they need to be merged into the existing clutter layer in a format the propagation tool can ingest.
Standard clutter layer formats for tools such as Atoll, EDX, and similar ray-tracing or empirical propagation engines use rasterised building height grids at 1 to 5 m pixel spacing, often accompanied by a vector building footprint layer with height attributes. The change polygons can be burned into the existing raster at the detected height, replacing only the cells that changed. This partial update approach preserves the investment in the original clutter dataset and avoids a full re-derivation. The propagation team can then re-run only the affected small-cell candidates, rather than the entire network plan.
One honest caveat: stereo DSMs do not resolve interior building geometry. A new building detected at 30 m height is attributed that height uniformly across its footprint. Actual floor-plate setbacks, podium levels, or irregular facades require either a LiDAR survey or manual editing. For most 5G small-cell planning purposes, the uniform-height approximation is acceptable; for millimetre-wave links where a 2 m parapet matters, it is not.
Practical workflow and what to order
A workable change-detection workflow for a city-scale small-cell re-assessment runs as follows. First, screen the candidate site list against a recent Sentinel-2 time series to identify which sites have construction activity within a 200 m radius. This costs nothing beyond processing time, since Sentinel-2 data is freely available through the Copernicus Data Space. Second, task Pleiades Neo or WorldView Legion stereo over the flagged sites only, which concentrates commercial imagery spend where it is needed. Third, generate baseline and current DSMs, compute the height-change raster, run the attribution classifier, and produce the updated clutter polygons. Fourth, deliver the polygons to the RF planning team in their preferred format for propagation re-runs.
Satellize runs this workflow on client licence, combining open Sentinel-2 screening with commercial stereo tasking. The Tonga crop-estimation programme demonstrated the same pattern of open-data screening followed by targeted commercial tasking, applied to a very different domain. The principle transfers directly: use free data to decide where to spend, then spend precisely.
Turnaround from tasking request to updated clutter layer delivery depends almost entirely on cloud-free acquisition, which is outside anyone's control. In temperate cities with reliable clear-sky windows, two to three weeks is a reasonable expectation. In humid tropical cities, build in a contingency of four to six weeks and consider whether a dry-season tasking campaign is more efficient than on-demand orders.
Typical figures
| Native panchromatic resolution | 30 cm (Pleiades Neo, WorldView Legion) |
| Stereo DSM vertical accuracy | 0.5–1 m RMSE with GCPs; 1–2 m RMSE without, in urban areas |
| Stereo DSM horizontal accuracy | Better than 1 m CE90 with rational polynomial coefficients |
| Minimum detectable height change | Approximately 3 m (single-storey) with caution; 6 m and above with high confidence |
| Screening layer revisit (Sentinel-2) | 5 days at equator (two-satellite constellation), 10 m resolution |
| Commercial stereo revisit | Up to twice daily (Pleiades Neo); under 30 minutes over major cities (WorldView Legion) |
| Archive depth | Pleiades/SPOT archive from 2011; Sentinel-2 from 2015; WorldView from 2007 |
| Cloud cover constraint | Both stereo passes must be cloud-free; tropical cities may require 4–6 week acquisition windows |
| Delivery formats | GeoTIFF height-change raster, GeoPackage or Shapefile change polygons with height and class attributes, CSV site-level summary |
| Clutter layer output resolution | 1–5 m raster (matches standard RF propagation tool inputs) |
Analytics Satellize can run
| Construction activity screening map | Sentinel-2 multitemporal spectral change detection (normalised difference built-up index and bare-soil spectral signatures across time series) | GeoTIFF and site-list CSV flagging candidate sites with detected construction activity within configurable radius |
| Baseline and current digital surface models | Semi-global matching or similar dense stereo reconstruction from Pleiades Neo or WorldView Legion stereo pairs | GeoTIFF DSMs at 0.5–1 m pixel spacing, one per epoch, georeferenced to WGS84 |
| Height-change raster | Pixel-wise DSM differencing with co-registration correction and outlier filtering | GeoTIFF signed height-change layer with per-pixel confidence band |
| Change-class attribution polygons | Multi-temporal stacking and morphological classification to separate new construction, demolition, temporary obstructions (cranes), and ambiguous change | Vector polygon layer (GeoPackage or Shapefile) with attributes: height delta, class label, confidence score, acquisition dates |
| Updated clutter layer patch | Rasterisation of change polygons burned into existing clutter raster at detected building height; partial update preserving unchanged cells | GeoTIFF clutter raster in client-specified resolution (1–5 m) ready for import into Atoll, EDX or equivalent propagation tool |
| Site-level re-assessment report | Automated per-site summary of detected obstructions within line-of-sight corridor, ranked by height-change magnitude | PDF and CSV report listing each candidate small-cell site, new obstructions detected, height and distance from site, recommended propagation re-run priority |
Who does the work
We can get this done for you. Satellize runs its own analyst desk and a strong science team. You do not buy a data feed and work out what it means; our people source the imagery, run the analysis described on this page, and hand you the answer with its confidence limits stated. Discuss this requirement.